Video: face jitters

Hi,

I've tested the video swap functionality using the provided Colab, and in every test, the face "jitters". It looks like the scale of the face is changing by a very small amount. Sometimes the eyebrows are not aligned, either. So we have two copies of the eyebrows (in the instance that the original video raises their eyebrows).

Also, the mouth & eyes do not correctly match the source.

This video shows an example of the jittery face & eyes not pointing correctly.

https://user-images.githubusercontent.com/87626308/151474362-137800dd-d38e-4180-99a2-20f9f546363d.mp4

And this shows how the mouth should look in the source, how SimSwap correctly translates it, and how sber-swap does not have the correct eyes or mouth.


Are these issues with the approach, or would training the model for longer resolve the issues?